rokosbasilisk comments on A central AI alignment problem: capabilities generalization, and the sharp left turn